Multi-label learning of non-equilibrium labels completion with mean shift

• Published in: Neurocomputing (Amsterdam) Vol. 321; pp. 92 - 102
• Main Authors: Yusheng, Cheng, Dawei, Zhao, Wenfa, Zhan, Yibin, Wang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 10.12.2018
• Subjects: Information entropy; Label completion; Label correlation; Mean shift; Multi-label classification; Mean shift; Label correlation; Multi-label classification; Information entropy; Label completion
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2018.09.033

In multi-label learning, the use of labels correlation is crucial for the improvement of multi-label learning performance. Most of the existing methods for studying labels correlation usually do not consider the study of feature-space information. Further study is deserved about how to synchronize rich information contained in features-space and labels-space. In this paper, a multi-label learning algorithm of Non-Equilibrium Labels Completion with Mean Shift (i.e. NeLC-MS) was proposed. The aim of this research was to mine the feature hidden information by reconstructing the features space, and introduce non-equilibrium label correlation information so as to better improve the robustness of multi-label learning classification. First, the mean shift clustering method was used to reconstruct the information between features in the feature space to obtain the hidden information between features. Then, the new information entropy was used to measure the correlation between labels which gets the basic labels confidence matrix. Then the basic labels confidence matrix was improved to construct a Non-equilibrium labels completion matrix by the non-equilibrium parameters. Finally, the new training set was constructed by using the reconstructed features space and the Non-equilibrium Labels Completion matrix, and the existing linear classifier was used for predicting the new training set. The experimental results of the proposed algorithm in the opening benchmark multi-label datasets showed that the NeLC-MS algorithm would have some advantages over other comparative multi-label learning algorithms, and the effectiveness of the proposed method was further illustrated by the use of statistical hypothesis test and stability analysis.